在 m 大小的窗口中查找最小 n 值的移动平均值

Question

I have data on the values of personal stocks like this:

UserId Stock Value    Time
1        APL  20  '2019-01-01'
1        MCR  40  '2019-01-01'
1        ADX  60  '2019-01-01'
3        AGL  10  '2019-01-01'
...

I have to group by users and for each stock x, I have to find the average of the value the 10 most valuable stocks in that user's 20 most recent stocks before stock x. Thus, I first group by the userId, then iterate through each stock x doing the following: select the user's 20 most recent stocks before stock x, further select the 10 most valuable stocks from that window, take the average and add it on in a new column for stock x. My dataset would look something like this after:

UserId Stock Value    Time    MovingAverage
1        APL  20  '2019-01-01'     20
1        MCR  40  '2019-01-01'     30
1        ADX  60  '2019-01-01'     40
3        AGL  10  '2019-01-01'     10
...

So far, I have been trying to use rolling in Python as follows:

df = df.sort_values(['userId','time'], ascending=['true','false']) 
df['roll'] = df.groupby('userId')['Value'].transform(lambda x: x.rolling(20,1).mean())

I can't figure out how to get the mean of the 10 highest values in a window! I am not against using another technique than rolling, it was just what seemed like the most popular method.

Another issue is that some stocks will have less than 20 stocks before them, but I think using rolling(20,1) mitigates that issue. However, in the case that there are less than 10 stocks eg 8 stocks, I need to just get the average of the last 8 stocks.

Answer 1

Figured it out. Posting in case anyone else is in a similar situation. I defined my own function and then simply used rolling.apply(). Ended up being fairly straightforward.

First, I defined the function that would perform the behaviour described in the post above.

def gm(arr):
    if (arr.size > 10):
        x = np.partition(arr, 9).mean()
    else:
        x = arr.mean()
    return x

Then, rolling.apply() worked its magic:

newcol = df.groupby('userId')['value'].rolling(20,1).apply(lambda x: gm(x), raw=True)
df['roll'] = newcol.reset_index(level=0, drop=True)

I am still not sure about the indexing at the end but the results seem to be what I want.